Research On Face Recognition Based On Weak Features

With the vigorous development of the Internet of Things related industries and the large-scale popularization of computer networks,modern identity authentication is developing in the direction of unmanned and intelligent,which objectively further promotes the rapid development of identity authentication based on face recognition.In the task of face recognition,face recognition has always been one of the main research contents of computer vision and deep learning,and accurate face detection,as an important technology in the face recognition process,has always been a research hotspot in related fields.Face detection refers to a technology or method by which a computer can accurately locate the position of a portrait from a photo with an uncertain position of the portrait.This paper carries out the research on face detection from three aspects.First,find a wealth of portrait data materials from the Internet.Considering that this article is based on the research of face recognition under weak features,the ms1 m and LFW datasets with unconstrained conditions are selected as the training set and the evaluation set,respectively,secondly,the common light processing methods in the image processing field are used to alleviate the influence of strong and weak light on the subsequent recognition rate;finally,the best effect in the current target detection field is used Deep learning network Retinanet,and determine the depth of the backbone network at 50 layers,introduce shortcuts to alleviate the degradation of the network,introduce the Bottleneck structure to enhance the network's nonlinear representation ability,and use Focol Loss to solve the Class Imbalance that affects the accuracy of one stage.The activation function in the network is changed to PRelu,and the landmark obtained by the face detection frame regression is subjected to affine transformation to achieve portrait alignment.In the face recognition process,Resnet with Backbone 101 is used,and the SE module is introduced for network-level nonlinear enhancement,and the European metric method is determined as the final comparison method of features.In the final face search link,Faiss vector encapsulation index to improve algorithm speed.In the realization of this article,Pytorch programming is used.The face recognition process of semantic segmentation is demonstrated using visualization.Finally,the performance indicators of the algorithm were evaluated in the matplotlib software,which verified the effectiveness of the algorithm from an experimental perspective;in addition,in the evaluation stage,through the comparison results of the recognition rate of this algorithm and other algorithms on the LFW and YTF data sets,it was verified The robustness of the algorithm is improved.